robert-dean/deadair v0.1.0

deadair 0.1.0

on GitHub

The first release. Everything below has been running on one station for some time; what is new is
that there is now a number to name it by.

The station

• A running order, owned by one thing. A forward lineup several hours deep, every item carrying
  its own state, with the director as its only writer — the console, the schedule and the model all
  post commands to it rather than writing it themselves.
• A presenter that cannot be silenced by a model. Breaks are written by a local or hosted model
  when one is configured, by the operator's own phrasings when none is, and by the station's own
  five underneath both. The floor cannot fail.
• Facts, or nothing. A claim the presenter states on air is a stored row carrying the sentence
  of source prose that supports it. A claim with no source is not expressible in the schema.
• Personas that accumulate. A character sheet, a voice, a notebook of what it has said and
  anecdotes it can tell, plus auditions and rehearsals that never reach air. Phone-ins are produced
  as a cast of personas trading turns, each its own model call, joined into one file before it airs.
• Measurement before air. A Python sidecar decodes each record and answers with its cue points
  and its loudness, so silence is trimmed and boundaries are sized from what the material does.
• An answer to "why is it quiet". Eleven ordered gates over one snapshot produce a single causal
  verdict, written to the station's own event log.
• A format clock and a weekly schedule, a catalog with the station's opinion of every record at
  three levels, and an activity feed of what actually happened.

Around it

• One container carrying the station, its console, the audio chain, the stream server and the
  sidecar, in three variants: slim, latest and full.
• Plugins for music providers (Spotify, Navidrome), enrichment (MusicBrainz, Last.fm,
  Wikipedia), news (RSS), web search (SearXNG, Brave, Tavily), weather (Open-Meteo, the US National
  Weather Service, OpenWeatherMap), speech (Kokoro, Chatterbox), models, and the measurement
  adapter.
• Listener apps for Android and macOS, released separately and on their own schedules.

Known limits

• Images are linux/amd64 only. There is no arm64 build yet.
• There is no MFA recovery code. A lost authenticator is recovered against the database; the
  procedure is in the README.
• latest follows main. Pin 0.1 to track releases only.